Review is 70–80% of discovery spend, and the driver is not volume — it is process, opaque billing units, and charges nobody quotes you. Five drivers the Winter 2026 pricing survey exposes, and what to do about each.
Ask most litigation teams why their last matter ran over budget and you will get a version of the same answer: too much data. So the response is more culling, tighter keywords, more aggressive deduplication. All sensible. None of it addresses the largest line item.
Document review accounts for an estimated 70–80% of total eDiscovery spend, and the driver of that cost is not how many documents reach review — it is how the review is run. eDiscovery document review costs are a process problem wearing a volume costume.
The Winter 2026 eDiscovery Pricing Survey, published by ComplexDiscovery OÜ in partnership with EDRM, gives us the first credible market-wide picture of what that process actually costs. Below are the five drivers the data exposes — and what legal operations teams can do about each.
The through-line: every driver below is a form of ambiguity — an unstated billing unit, an uncontrolled collection scope, an unwritten review protocol, an unpriced exception clause, an unknown rate. Cost does not enter through volume. It enters through the gaps in what was agreed.
Scroll table horizontally →
| # | Driver | What it actually costs you |
|---|---|---|
| 1 | The billing unit | Modelling a matter against a pricing structure your provider no longer uses |
| 2 | Over-collection | Hosting fees on data that should never have been promoted — every month, for the life of the matter |
| 3 | The process tax | Second and third passes over the same population, invoiced as review hours rather than rework |
| 4 | Unquoted charges | Exception documents and future token repricing, neither of which appears in the quote |
| 5 | The transparency gap | An AI business case you cannot evaluate because you never knew the human baseline |
Discovery pricing has quietly migrated from user and licence arrangements toward data-volume models — for example, a flat per-GB rate. That shift is rational for providers, but it changes where cost risk sits, and a lot of firms are still pricing matters against the old framework.
The Winter 2026 numbers show a market split cleanly in two. Commodity infrastructure has bottomed out: 54.7% of respondents report basic hosting below $10/GB/month, and processing at ingestion sits below $75/GB for 73.6%. But the differentiated layers above that infrastructure hold their price — 11.3% report analytics-enabled hosting above $25/GB/month.
The more telling signal is how many respondents cannot name a unit at all. Alternative pricing models account for 34.0% of user licensing responses and 35.8% of predictive coding and TAR responses — the highest in the survey. Per-GB TAR billing is being absorbed into platform subscriptions, which is convenient right up until you try to model a matter.
What to do: establish, at engagement, whether your provider bills an all-in per-GB rate covering ingestion through promotion, or a staged model with separate charges at each step. The survey’s own authors flag this as a distinction practitioners routinely miss. Our breakdown of per-GB hosting fees covers what each variant typically includes.
Volume is not the whole story, but it is not nothing — and most of it is self-inflicted at the front end.
A useful illustration: a standard email without attachments, printed to PDF, came out at 126 KB. The same email in native .pst was 29 KB. You could store four native emails in the space one PDF occupies — before considering the integrity and authenticity problems the conversion introduces. Multiply that across a 500 GB collection and the format decision alone moves the hosting bill by a factor of four, every month, for the life of the matter.
Three techniques do most of the work here: targeted collection instead of accepting data dumps, modern collection tooling, and advanced processing that culls before promotion rather than after. The distinction matters more than it sounds. Culling after promotion reduces what reviewers see; culling before promotion reduces what you are billed for.
This is the driver that never appears as a line item.
Inconsistent coding decisions. Re-review when the protocol shifts mid-matter. Partners reconciling conflicting calls between reviewers. Reconciliation passes that exist purely because the first pass was not reproducible. None of it is billed as “rework.” All of it lands in the final invoice as more review hours.
The survey data makes the arithmetic concrete. Managed review attorney rates exceed $40/hour for 45.3% of respondents onsite and 35.8% remote, and per-document rates cluster in the $0.50–$1.00 range for both delivery models. Every avoidable second pass over a 500,000-document population is a six-figure event.
And project management — the function that is supposed to prevent exactly this — is getting more expensive as engagements get more complex: 26.4% of respondents now report PM rates above $200/hour. Paying premium rates for coordination is a reasonable trade only if the coordination is actually reducing rework.
Two findings in the survey should worry anyone signing a GenAI-assisted review agreement.
First, exception documents — the ones that fail AI processing or need human intervention. Nearly 40% of respondents (39.6%) cannot say how their contracts handle them. Among those who can, there is no standard:
In a matter where a meaningful share of documents exits the AI workflow, that clause is the difference between an on-budget review and a blown one.
Second, per-token pricing has barely reached buyers — only 5.7% report it as their primary model. Providers are currently absorbing LLM cost variability inside bundled rates. One survey respondent cautioned that token-based pricing pressure could push AI costs up materially in future absent significant reductions in GPU infrastructure costs. If you are signing a multi-year commitment, ask what happens when that absorption stops — and whether the platform is model-agnostic enough that the exposure is yours to manage rather than your vendor’s to reprice.
You cannot optimise what you cannot see, and a striking share of the market cannot see its own review rates.
For onsite per-document review, 34.0% of respondents do not know the cost. For remote, 30.2%. These are not outsiders — they are practitioners buying review services. Nearly one in five also cannot name the hourly attorney rate underlying their managed review engagements (18.9% onsite, 17.0% remote).
That gap has a direct commercial consequence. Human per-document review clusters at $0.50–$1.00. GenAI-assisted review clusters at $0.11–$0.50, with the $0.26–$0.50 tier the single most-cited price point. The economic case for AI-assisted review is legible only if you know your human baseline. A third of the market does not.
For in-house teams this matters more than for firms. Law firms pass eDiscovery costs through to clients; corporate legal departments absorb them entirely — which makes document review pricing transparency a budget integrity issue rather than a negotiation tactic.
The pattern across all five drivers is the same: cost enters through ambiguity. Ambiguous billing units, ambiguous collection scope, ambiguous review protocols, ambiguous exception terms, ambiguous rates. Reducing eDiscovery document review costs means closing each of those in turn.
DecoverAI is priced against exactly this problem: $60/GB/month, flat and all-in. AI document classification and review, auto-generated privilege logs, Bates numbering, redactions, semantic search, and full production delivery are inside that number. There are no seat fees — reviewers access a matter without adding to the bill — and no contracts or minimums.
Two things follow from that structure, and they map onto the drivers above. Because every AI capability is bundled rather than metered as credits, there is no usage to forecast and no second invoice arriving after the fact — the token-cost variability that survey respondents flagged as a forward risk does not sit on your side of the line. And because the rate covers the review workflow end to end rather than staging charges at ingestion, promotion, production, and privilege log, the effective per-document cost on a matter is arithmetic you can do before you start rather than a number you reconstruct from invoices afterward.
Most legal teams process 5–20 GB per matter, which puts a typical matter in the $300–$1,200 range all-in. The cost estimator will model your own volumes, and the pricing page lists what is inside the rate line by line.
The Winter 2026 survey drew 53 responses, 92.5% from the United States, and reports self-reported practitioner perceptions of prevailing market pricing rather than verified transaction records. Treat the figures as directional market intelligence — useful as a negotiation baseline, not as an audited benchmark. The 70–80% review share is a practitioner estimate and is not sourced from the survey data.
What percentage of eDiscovery cost is document review?
Practitioner estimates put document review at 70–80% of total eDiscovery spend. That share is not primarily a function of how many documents reach review. It is a function of how the review is run — how many passes are made over the same population, how consistent the coding is, and how much of the work is rework that never appears on an invoice as rework.
What does document review cost per document?
In the Winter 2026 survey, human per-document review clusters in the $0.50–$1.00 range for both onsite and remote delivery. GenAI-assisted review clusters at $0.11–$0.50, with $0.26–$0.50 the single most-cited tier. These are self-reported perceptions of market pricing rather than verified transactions, so use them as a negotiation baseline rather than an audited benchmark.
What are exception documents, and how are they billed?
Exception documents are files that fail AI or automated processing and require human intervention — corrupt files, unsupported formats, encrypted containers, and documents the model cannot classify with confidence. There is no market standard for billing them: 39.6% of survey respondents could not say how their contracts treat exceptions at all, and among those who could, 18.9% route them to manual review at standard rates, 17.0% say it depends on the issue, 9.4% see additional processing charges, and 9.4% report no additional charge. Get the treatment written into the agreement before signing.
How do I reduce eDiscovery review costs?
Close the ambiguities that let cost in. Cull before you host rather than after. Make the first pass the only pass, using consistent AI-assisted relevance and privilege calls with auditable reviewer decisions — rework is the real cost centre. Price exception documents explicitly in the contract. Ask for an effective per-document rate even when your provider bills hourly. And know whether you are paying for infrastructure, which has commoditised, or for analytics and AI, which have not.
Is per-token pricing common in eDiscovery contracts?
Not yet — only 5.7% of respondents report it as their primary model, which means providers are currently absorbing LLM cost variability inside bundled rates. One respondent cautioned that token pricing pressure could push AI costs up materially in future absent significant reductions in GPU infrastructure costs. If you are signing a multi-year commitment, ask in writing what happens to your rate when that absorption stops.
Why can’t I compare AI-assisted review to human review?
Usually because you do not know your own human baseline. A third of the market cannot name its per-document review cost, and roughly one in five cannot name the hourly attorney rate underlying its managed review engagements. The economic case for AI-assisted review is legible only against a known human number, so establishing that number is the first step — before the comparison, not after it.
This article is general information about legal technology and discovery practice, not legal advice for any particular matter. Third-party pricing figures reflect self-reported survey responses collected in Winter 2026 and may not describe current market rates or the terms available to you; verify current pricing directly with each vendor. DecoverAI pricing described here reflects published rates at the date of writing and is subject to change.